Artificial Intelligence, Personalised Persuasion, and Climate Attitudes
European Institute, London School of Economics
Climate Information and Climate Concern
- The Problem: Persuading people to care about and act toward climate change is still one of the major collective challenges we face as a society
- Effective communication is crucial
- Over $15 million spent in science-based public information campaigns (United Nations 2025)
- However, effects are small and short-lived (Deryugina and Shurchkov 2016; van der Linden 2017).
In Practice:
Why is this happening?
- The issue: Traditional information campaigns often fail because they don’t engage people personally
- Psychological distance and intertemporal discounting reduce climate urgency (Trope & Liberman, 2010).
- Emotional disconnection from abstract risks (Weber, 2006).
The Problem
People see why climate matters generally, but not why it matters to them
The solution:
- Tailored messaging aligned with individual beliefs and experiences is more effective
- But traditional targeting faces critical limitations:
- Operates at group level (demographics, ideology)
- Expensive
- ⟹ Hard to scale
AI as a Solution
Key Advantages:
- From a practical perspective, AI can:
- Scale: Reach millions with personalised messages
- Cost-effective: Reduce costs significantly
- Engage: Create tailored conversations to the individual level
- From a theoretical perspective, AI can:
- Evidence-based reasoning (Costello et al., 2025)
- Non-judgmental interaction (Goel et al., 2024)
- Responsive dialogue (Santoro et al., 2024)
Hypotheses
H1: When compared with the control condition (information provision), interacting with the LLM will have a positive effect on all outcomes of interest: a) climate concern, b) policy support, c) pro-environmental behavior, d) conversation spillover, e) political spillover, and f) subjective climate beliefs.
H2: The personalized LLM interaction on climate change will be more persuasive/effective than the non-personalized LLM interaction and the basic information provision on all outcomes of interest.
H3: The personalized LLM interaction on climate change condition will be more persuasive/effective compared to the personalized AI interaction on an unrelated political topic.
Research Design
Online Survey Experiment:
- Sample
- N = 1,500-2,000 UK adults (representative sample)
- Recruited via established survey platform (Prolific)
- Power analysis: 930 minimum for small-to-medium effect (d = 0.26-0.37)
- Context: UK offers valuable setting where climate skepticism is growing but not deeply entrenched
The treatments
Participants are randomized into these conditions:
Control: Standard information provision about climate change
Non-Personalized AI: Generic LLM conversation about climate change
Personalized AI (Unrelated Topic): Personalised conversation about a political topic unrelated to climate change
Personalized AI (Climate): Personalised climate conversation about climate change:
Survey Flow
An Example
What this looks like?
Outcome Variables
- Climate concern: 5-point scale from “Not at all concerned” to “Extremely concerned”
- Pro-environmental behavior: Willingness to donate portion of $100 prize to climate NGO
- Conversation spillover: Likelihood of discussing AI conversation with others
- Political spillover: Choice to write message to Member of Parliament about climate change
- Subjective climate beliefs: Open-ended question about personal beliefs on climate change + LLM-generated Likert scale
Design Advantage
Pre-post measurement enables both within-subject change detection and between-group comparisons
Field Experiment Integration
Social Media
Chat with our AI for 5 minutes
Newsletter
• Local community wins against fracking proposal
• New study shows impact of diet on carbon footprint
The Numbers
Thank you!
Any Questions?